An Adjustable Feature-Weighted Bayesian Model for Hybrid Satellite Telemetry Variables Anomaly Detection Under Multioperating Conditions

Integrated health management of the main node satellite is crucial to ensuring overall safety and stability of the large-scale constellation. Nevertheless, the telemetry data of the main node satellite includes several categories of characteristics, and the satellites may encounter a series of conditions of slow or abrupt transition, all of which provide a considerable challenge to the satellite anomaly detection task. Traditional machine learning methods such as KNN, SVM, and PCA cannot effectively utilize multi-valued variables, while deep learning methods generally lack interpretability. As a result, based on the Bayesian theory and statistical learning methods, this paper proposes an adjustable feature weighted Bayesian model (AFWBM) that can utilize both continuous and multi-valued variables and is completely interpretable. AFWBM employs the Gaussian mixed and multinomial mixed distributions to describe continuous and multi-valued variables under multi-operating conditions, respectively. Additionally, for the initial label formulation and parameters learning tasks, the expectation-maximization (EM) algorithm is utilized. Theoretically, AFWBM computes the weights of each feature using normalized mutual information (NMI) and takes the label of the maximum prediction probability as the prediction label. Since most faults in the satellite anomaly detection tasks are usually accumulated from the aging of health components, AFWBM updates the initial labels based on the rapid aging process determination scheme. Furthermore, the incremental learning method is adopted for online adjustment of model weights and parameters, as well as online reprediction of sample labels. Finally, a numerical example and a practical example of satellite are utilized to verify the superiority of AFWBM under both dual-working and multi-working anomaly detection tasks.